Greenhouse and protected agriculture structure mapping
Plastic greenhouses and polytunnels produce a distinctive high-reflectance signature in visible and SWIR bands that saturates standard vegetation indices. Detecting them accurately requires resolving the confusion with salt pans, rooftops, and sand, and knowing where optical sensors alone fall short.
Sensors
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in SWIR. Five-day revisit at mid-latitudes. The SWIR bands (1610 nm and 2190 nm) are particularly diagnostic: polyethylene reflects strongly where healthy vegetation absorbs, producing a clear spectral inversion. Free and globally archived from 2015.
- Planet SuperDove: 3 m resolution, eight spectral bands including red-edge and NIR. Near-daily revisit over most land surfaces. Sufficient to resolve individual polytunnel bays and detect small structures below Sentinel-2's practical detection threshold of roughly 200–400 m². Requires a commercial licence.
- Maxar WorldView-2 / WorldView-3: 30–46 cm panchromatic, 1.2–1.85 m multispectral. Resolves individual tunnel ribs and frame shadows. WorldView-3 adds eight SWIR bands at 3.7 m, enabling direct spectral fingerprinting of cover materials. Revisit is 1–4 days depending on tasking priority and latitude. High cost per km² limits use to targeted verification.
- Sentinel-1 SAR (C-band): 10–20 m resolution in Interferometric Wide Swath mode, six-day repeat at mid-latitudes (three days with both satellites). C-band backscatter distinguishes metallic greenhouse frames, which produce strong double-bounce returns, from plastic polytunnels, which scatter more diffusely. Cloud-independent, so it fills optical gaps during overcast periods common in winter growing seasons.
What a polyethylene roof gives away
Healthy vegetation absorbs strongly in the red and blue bands and reflects in the near-infrared, producing the familiar high NDVI signature. Polyethylene does the opposite. It reflects across the visible spectrum and into SWIR, producing NDVI values close to zero or slightly negative, and SWIR reflectance values that rival bare sand. In Sentinel-2 Band 11 (1610 nm), a dense greenhouse complex appears almost as bright as a salt flat.
This spectral inversion is the foundation of every automated greenhouse-mapping method in the published literature. Indices designed to isolate built surfaces, such as the Normalised Difference Built-up Index (NDBI), partially capture greenhouse clusters, but purpose-built indices that combine SWIR brightness with low NIR absorptance perform better in practice. The confusion risk is real: glass greenhouses have a slightly different spectral profile from polyethylene because glass absorbs more in SWIR, which matters when you are trying to distinguish heated glasshouses from unheated polytunnels.
Resolution floors and the small-tunnel problem
A standard Spanish-style polytunnel is roughly 8 m wide and 50–100 m long. At Sentinel-2's 10 m pixel size, a single tunnel occupies perhaps one pixel in width. Reliable area estimation at that scale requires that clusters of tunnels cover at least four to six contiguous pixels, which in practice means a minimum detectable cluster of roughly 400–600 m². Individual tunnels and small family-scale structures are effectively invisible to Sentinel-2 unless they aggregate into a larger complex.
Planet SuperDove at 3 m resolves individual tunnel bays clearly. WorldView at sub-metre panchromatic resolution resolves structural elements such as ventilation gaps and access paths, which is useful for verifying material type but not cost-effective for regional surveys. The practical workflow for most government clients is a two-stage approach: Sentinel-2 for national or regional extent mapping, then Planet or Maxar tasking for ambiguous parcels and change verification.
Confusion sources that will fool a naive classifier
Salt pans, dry lake beds, and sandy beaches produce SWIR reflectance profiles that overlap significantly with polyethylene. Urban rooftops, particularly metal-sheeted industrial buildings, add further confusion. A classifier trained only on spectral features will misidentify all of these as greenhouse cover in arid or coastal regions.
Several published approaches address this. Temporal consistency is the most reliable discriminator: greenhouse structures persist across seasons while salt and sand surfaces show no regular geometric pattern and vary with rainfall and tide. Geometric texture features, such as the regular linear arrangement of tunnel ridges visible in high-resolution imagery, provide a second filter. SAR backscatter adds a third dimension: metallic roofs produce strong double-bounce returns in Sentinel-1 VV polarisation, whereas polyethylene scatters more diffusely and produces lower, more uniform backscatter. Combining optical and SAR features in a random-forest or gradient-boosting classifier substantially reduces false positives in coastal agricultural zones.
SAR as a cloud-independent cross-check
Mediterranean and East Asian greenhouse districts, where the majority of the world's plastic-covered agriculture is concentrated, frequently experience winter cloud cover precisely when crop cycles are most active. Sentinel-1's C-band SAR penetrates cloud and acquires consistently regardless of illumination, making it the practical solution for monitoring during those periods.
The backscatter mechanism differs by material. Metallic greenhouse frames oriented perpendicular to the radar look direction produce a strong double-bounce return, elevating sigma-naught in VV polarisation by several decibels relative to surrounding bare soil. Plastic tunnels, lacking metallic elements, return a weaker and more diffuse signal. This difference is exploitable: a SAR-only classifier can separate metallic glasshouse complexes from polytunnel clusters with reasonable accuracy in flat terrain. In hilly terrain, slope-induced radiometric distortion complicates interpretation, and SAR should be used alongside optical data rather than alone.
Building a change-detection programme, not a one-time map
A single-epoch greenhouse map is of limited policy value. Expansion of plastic agriculture in southern Spain, Morocco, China's Shandong province, and parts of sub-Saharan Africa has been rapid and largely unregistered. Governments and water-resource managers need annual or seasonal updates to track encroachment, estimate water demand, and enforce land-use regulations.
Dense time-series analysis using the full Sentinel-2 archive, which extends to 2015, allows retrospective reconstruction of expansion rates. Change detection between annual composites, using the SWIR-based greenhouse index as the primary signal, identifies new structures to within one growing season. The Satellize analytics pipeline applies this approach operationally; the methodology is directly analogous to the crop-area estimation work done for the Kingdom of Tonga, adapted for the spectral signature of covered rather than open agriculture. Honest caveat: in regions where cloud frequency exceeds 60% of annual observations, annual composites may still contain gaps that require SAR gap-filling or interpolation, introducing uncertainty in the change estimates.
What the map cannot tell you on its own
Area extent is not the same as production. A greenhouse map tells you where covered agriculture exists and how it is changing. It does not tell you what crop is inside, whether the structure is actively heated, or what yield is being achieved. Combining the structure map with thermal infrared data can indicate which glasshouses are heated, since operational heated greenhouses show elevated land-surface temperature at night relative to unheated polytunnels. But thermal resolution from Landsat (100 m) and ECOSTRESS (70 m) is coarse relative to individual tunnel dimensions, so this inference works best for large glasshouse complexes.
Material classification from spectroscopy is improving. WorldView-3's SWIR bands can distinguish polyethylene from polypropylene and glass in controlled conditions, but atmospheric correction quality and within-scene variability mean operational material classification remains difficult. Treat material-type attribution as indicative rather than definitive unless supported by ground-truth sampling.
Typical figures
| Typical spatial resolution (optical) | 10 m (Sentinel-2), 3 m (Planet SuperDove), 0.3–1.85 m (Maxar WorldView) |
| Typical spatial resolution (SAR) | 10–20 m (Sentinel-1 IW mode) |
| Revisit frequency | 5 days (Sentinel-2, mid-latitudes); near-daily (Planet); 6 days / 3 days with both satellites (Sentinel-1); 1–4 days on tasking (Maxar) |
| Key spectral bands | SWIR 1610 nm and 2190 nm (polyethylene signature); Red and NIR (vegetation index inversion); C-band 5.4 GHz (SAR backscatter) |
| Minimum detectable cluster (Sentinel-2) | Approximately 400–600 m² (4–6 contiguous pixels); individual tunnels below this threshold require Planet or Maxar |
| Archive depth | Sentinel-2 from 2015; Landsat back to 1972 (30 m, limited SWIR utility for small structures); Planet from approximately 2016 depending on region |
| Cloud limitation | Optical sensors blocked by cloud; SAR (Sentinel-1) cloud-independent. Dense cloud regions require multi-source compositing |
| Mapping latency (operational) | 2–5 days from acquisition for automated change alerts; 2–4 weeks for validated annual extent maps |
| Typical delivery formats | GeoTIFF extent rasters, GeoJSON or Shapefile polygon layers, area-statistics tables by administrative unit, seasonal change reports |
| Coverage | Global; Sentinel-2 covers all land surfaces between 84°N and 56°S |
Analytics Satellize can run
| National greenhouse extent map | SWIR-based spectral index thresholding combined with geometric texture filtering on Sentinel-2 annual composites | Polygon GIS layer with area statistics by administrative unit, annual update cadence |
| Year-on-year expansion report | Bi-temporal change detection on Sentinel-2 SWIR composites, with SAR gap-filling for high-cloud regions | PDF report with mapped change polygons and tabulated area gains and losses by region |
| Material-type classification (metallic vs. plastic) | Fusion of Sentinel-1 VV backscatter and Sentinel-2 SWIR reflectance in a random-forest classifier | Classified raster layer distinguishing metallic glasshouses from plastic polytunnels, with confidence scores |
| Small-structure verification layer | Object-based image analysis on Planet SuperDove or Maxar imagery for parcels flagged as ambiguous by regional mapping | Verified polygon layer with individual tunnel delineation for targeted parcels |
| Seasonal construction and removal alerts | Dense time-series monitoring on Planet near-daily stack; threshold breach triggers alert | Automated alert feed (GeoJSON or email) when new structures exceed a defined area threshold within a monitored zone |
| Water-demand proxy estimate | Greenhouse area multiplied by published crop-water-use coefficients for covered horticulture (FAO-56 method), disaggregated by administrative unit | Tabular report of estimated irrigation water demand by season, suitable for water-resource planning |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.